Local Interpretable Model-agnostic Explanations for models trained to predict players' wages in the FIFA 23 game¶
We analyse the random forest model trained in previous assignments on the FIFA 23 game dataset. The model aims to predict players' wages given their statistics in the game.
We use Interpretable Model-agnostic to analyse feature importance.
We analyse stability of the method.
We compare the LIME method with the SHAP method implemented during the previous assignment.
We also compare LIME explanations for the random forest and the linear model.
Analysys of stability of the LIME method¶
We compare LIME explanations for three observations and for each use five different seeds.
We observe that the method is quite stable - the most important features are usually the same across the seeds (note that Top 2 is the same for all observations), there are some differences in less important features.
Comparison of the first chosen observation¶
The first seed:
The second seed:
The third seed:
The forth seed:
The fifth seed:
Comparison of the second chosen observation¶
The first seed:
The second seed:
The third seed:
The forth seed:
The fifth seed:
Comparison of the third chosen observation¶
The first seed:
The second seed:
The third seed:
The forth seed:
The fifth seed:
Comparison betweem LIME explanations and SHAP explanations¶
We observe that explanations of LIME and SHAP methods may differ even for the most important feautures (see comparison of the second chosen observation). Also Top 3 features are consistent only for one out of five observations (the forth chosen observation).
However, comparing the sets of Top 3 most important features, we observe that at least two elements overlap for all 5 observations. This suggest that both methods are useful in separating important and unimportant features.
Comparison of the first chosen observation¶
LIME explanations:
SHAP explanations:
Comparison of the second chosen observation¶
LIME explanations:
SHAP explanations:
Comparison of the third chosen observation¶
LIME explanations:
SHAP explanations:
Comparison of the forth chosen observation¶
LIME explanations:
SHAP explanations:
Comparison of the fifth chosen observation¶
LIME explanations:
SHAP explanations:
Comparison of LIME explanations between the tree model and the linear model¶
The most important features for the random forest model and the linear model usually differ. The tree model mostly focuses on Overall and Value in euro features, while the linear model often focuses on Stats and Position ratings.
We observe that the most important feautures are usually consistent between lime and dalex library. There are some differences in less important features.
Comparison of the first chosen observation:¶
The random forest explanation with dalex library:
The random forest explanation with lime library:
The linear model explanation with dalex library:
The linear model explanation with lime library:
Comparison of the second chosen observation:¶
The random forest explanation with dalex library:
The random forest explanation with lime library:
The linear model explanation with dalex library:
The linear model explanation with lime library:
Comparison of the third chosen observation:¶
The random forest explanation with dalex library:
The random forest explanation with lime library:
The linear model explanation with dalex library:
The linear model explanation with lime library:
Comparison of the forth chosen observation:¶
The random forest explanation with dalex library:
The random forest explanation with lime library:
The linear model explanation with dalex library:
The linear model explanation with lime library:
Comparison of the fifth chosen observation:¶
The random forest explanation with dalex library:
The random forest explanation with lime library:
The linear model explanation with dalex library:
The linear model explanation with lime library:
Appendix¶
Importing libraries and dataset¶
# 1. Import libraries
!pip3 install shap
!pip3 install dalex
!pip3 install lime
!pip3 install -q condacolab
import condacolab
condacolab.install()
!conda install -c conda-forge python-kaleido
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import plotly
import kaleido
import pickle
import lime
import shap
import dalex as dx
from math import isclose
from sklearn.linear_model import LinearRegression
from sklearn.ensemble import RandomForestRegressor
# 2. Load dataset and models from the previous homework
with open('X_train.pickle', 'rb') as handle:
X_train_load = pickle.load(handle)
with open('y_train.pickle', 'rb') as handle:
y_train_load = pickle.load(handle)
with open('X_test.pickle', 'rb') as handle:
X_test_load = pickle.load(handle)
with open('y_test.pickle', 'rb') as handle:
y_test_load = pickle.load(handle)
with open('tree_model.pickle', 'rb') as handle:
forest_reg_load = pickle.load(handle)
with open('linear_model.pickle', 'rb') as handle:
linear_model_load = pickle.load(handle)
print(X_train_load)
print(y_train_load)
print(X_test_load)
print(y_train_load)
print(forest_reg_load.predict(X_train_load))
print(linear_model_load.predict(X_train_load))
print(forest_reg_load.predict(X_test_load))
print(linear_model_load.predict(X_test_load))
Analysys of the model¶
# 1. Observe predictions of 5 observations (POINT 1)
observations = X_test_load.sample(5, random_state = 1)
predictions_forest = forest_reg_load.predict(observations)
predictions_linear = linear_model_load.predict(observations)
print(observations)
print(predictions_forest)
print(predictions_linear)
# 2. Calculate lime decomposition for selected observations with lime library for the forest model (POINT 2, 5)
lime_explainer = lime.lime_tabular.LimeTabularExplainer(
training_data=X_train_load.values,
feature_names=X_train_load.columns,
mode="regression"
)
plot_id = 1
for i in range(len(observations)):
lime_explanation_forest = lime_explainer.explain_instance(
data_row=observations.iloc[i],
predict_fn=lambda d: forest_reg_load.predict(d)
)
_ = lime_explanation_forest.as_pyplot_figure()
plt.savefig('Point2_lime_forest_im' + str(plot_id) + '.png', dpi=300, bbox_inches='tight')
plot_id += 1
# 3. Calculate lime decomposition for selected observations with lime library for the forest model (POINT 2, 5)
plot_id = 1
for i in range(len(observations)):
lime_explanation_linear = lime_explainer.explain_instance(
data_row=observations.iloc[i],
predict_fn=lambda d: linear_model_load.predict(d)
)
_ = lime_explanation_linear.as_pyplot_figure()
plt.savefig('Point2_lime_linear_im' + str(plot_id) + '.png', dpi=300, bbox_inches='tight')
plot_id += 1
# 4. Calculate lime decomposition for selected observations with dalex library for the forest model (POINT 2, 5)
explainer_forest_dx = dx.Explainer(forest_reg_load, X_train_load, y_train_load, verbose = False)
plot_id = 1
for i in range(len(observations)):
explanation_forest_dx = explainer_forest_dx.predict_surrogate(observations.iloc[i])
explanation_forest_dx.plot()
plt.savefig('Point2_dalex_forest_im' + str(plot_id) + '.png', dpi=300, bbox_inches='tight')
plot_id += 1
# 5. Calculate lime decomposition for selected observations with dalex library for the linear model (POINT 2, 5)
explainer_linear_dx = dx.Explainer(linear_model_load, X_train_load, y_train_load, verbose = False)
plot_id = 1
for i in range(len(observations)):
explanation_linear_dx = explainer_linear_dx.predict_surrogate(observations.iloc[i])
explanation_linear_dx.plot()
plt.savefig('Point2_dalex_linear_im' + str(plot_id) + '.png', dpi=300, bbox_inches='tight')
plot_id += 1
# 6. Calculate Shapley values for selected observations with shap library (POINT 4)
explainer_shap = shap.TreeExplainer(forest_reg_load)
shap_values = explainer_shap.shap_values(observations)
plot_id = 1
for i in range(len(observations)):
shap.bar_plot(shap_values[i], feature_names = observations.columns, show = False)
plt.savefig('Point4_shap_im' + str(plot_id) + '.png', dpi=300, bbox_inches='tight')
plot_id += 1
# 7. Analyze stability of the method
plot_id = 1
for i in range(3):
for seed in range(5):
np.random.seed(seed)
explanation_forest_dx = explainer_forest_dx.predict_surrogate(observations.iloc[i])
explanation_forest_dx.plot()
plt.savefig('Point3_seed' + str(seed) + '_im' + str(plot_id) + '.png', dpi=300, bbox_inches='tight')
plot_id += 1
